US9521264B2

US9521264B2 - Echo removal

Info

Publication number: US9521264B2
Application number: US14/012,786
Authority: US
Inventors: Per Ahgren
Original assignee: Microsoft Technology Licensing LLC
Current assignee: Microsoft Technology Licensing LLC
Priority date: 2013-05-31
Filing date: 2013-08-28
Publication date: 2016-12-13
Also published as: WO2014194009A1; US20140357324A1; CN105379239A; EP2982102B1; CN105379239B; KR20160016880A; EP2982102A1; GB201309779D0; KR102111185B1; ES2613492T3

Abstract

Echo removal techniques are described. As part of the echo removal, an adaptive model estimate of the echo in a received audio signal is determined using an adaptive model based on an outputted audio signal and the received audio signal. The adaptive model executes an algorithm comprising a convergence parameter to determine filter coefficients and uses said filter coefficients to filter the outputted audio signal to determine the adaptive model estimate of the echo. An accuracy value of the adaptive model is determined according to an echo return loss enhancement metric. The convergence parameter is updated based on the accuracy value. The adaptive model estimate of the echo is used to remove the echo in the received audio signal.

Description

RELATED APPLICATION

This application claims priority under 35 USC 119 or 365 to Great Britain Application No. 1309779.5 filed May 31, 2013, the disclosure of which is incorporate in its entirety.

BACKGROUND

A device may have audio input apparatus that can be used to receive audio signals from the surrounding environment. The device may also have audio output apparatus that can be used to output audio signals to the surrounding environment. For example, a device may have one or more speakers for outputting audio signals and one or more microphones for receiving audio signals. Audio signals which are output from the speaker(s) of the device may be received as “echo” in the audio signal received by the microphone(s). It may be the case that this echo is not desired in the received audio signal. For example, the device may be a user device (such as a mobile phone, tablet, laptop, PC, etc) which is used in a communication event, such as an audio or video call, with another user device over a network. Far-end signals of the call may be output from the speaker at the user device and may be received as echo in the audio signals received by the microphone at the device. Such echo can be disturbing to users of the call, and the perceived quality of the call may be reduced due to the echo. In particular, the echo may cause interference for near-end audio signals which are intended to be received by the microphone and transmitted to the far-end in the call. Therefore echo cancellation and/or echo suppression may be applied to the received audio signals to thereby suppress the echo in the received audio signal. The power of the echo in the received audio signal may vary depending upon the arrangement of the user device. For example, the user device may be a mobile phone and in that case, the power of the echo in the received audio signal would normally be higher when the mobile phone is operating in a “hands-free” mode compared to when the mobile phone is not operating in a “hands-free” mode.

Echo cancellation (or “echo subtraction”) techniques aim to estimate an echo signal included in the audio signal received at the microphone, based on knowledge of the audio signal which is output from the speaker. The estimate of the echo signal can then be subtracted from the received audio signal thereby removing at least some of the echo from the received audio signal. Echo suppression is used to apply frequency-dependent suppression to the received audio signal to thereby suppress the echo in the received audio signal.

SUMMARY

This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

There is provided a method of removing echo in a received audio signal. As part of the echo removal, an adaptive model estimate of the echo in the received audio signal is determined using an adaptive model based on an outputted audio signal and the received audio signal. The adaptive model executes an algorithm comprising a convergence parameter to determine filter coefficients and uses said filter coefficients to filter the outputted audio signal to determine the adaptive model estimate of the echo. An accuracy value of the adaptive model is determined according to an echo return loss enhancement metric. The convergence parameter is updated based on the accuracy value. The adaptive model estimate of the echo is used to remove the echo in the received audio signal.

The method may be used in a call (e.g. a call implementing voice over internet protocol (VoIP) to transmit audio data between user devices) in which case the outputted audio signal may be a far-end signal received from the far-end of the call, and the received signal includes the resulting echo and a near-end signal for transmission to the far-end of the call.

BRIEF DESCRIPTION OF THE DRAWINGS

For a better understanding of the described embodiments and to show how the same may be put into effect, reference will now be made, by way of example, to the following drawings in which:

FIG. 1 shows a schematic illustration of a communication system;

FIG. 2 is a schematic block diagram of a user device;

FIG. 3 is a functional diagram showing modules of a user device for use in echo removal; and

FIG. 4 is a flow chart for a process of removing echo.

DETAILED DESCRIPTION

Embodiments will now be described by way of example only.

FIG. 1 shows a communication system 100 comprising a first user 102 (“User A”) who is associated with a first user device 104 and a second user 108 (“User B”) who is associated with a second user device 110. In other embodiments the communication system 100 may comprise any number of users and associated user devices. The

user devices

104 and 110 can communicate over the network 106 in the communication system 100, thereby allowing the

users

102 and 108 to communicate with each other over the network 106. The communication system 100 shown in FIG. 1 is a packet-based communication system, but other types of communication system could be used. The network 106 may, for example, be the Internet. Each of the

user devices

104 and 110 may be, for example, a mobile phone, a tablet, a laptop, a personal computer (“PC”) (including, for example, Windows™, Mac OS™ and Linux™ PCs), a gaming device, a television, a personal digital assistant (“PDA”) or other embedded device able to connect to the network 106. The user device 104 is arranged to receive information from and output information to the user 102 of the user device 104. The user device 104 comprises output means such as a display and speakers. The user device 104 also comprises input means such as a keypad, a touch-screen, a microphone for receiving audio signals and/or a camera for capturing images of a video signal. The user device 104 is connected to the network 106.

The user device 104 executes an instance of a communication client, provided by a software provider associated with the communication system 100. The communication client is a software program executed on a local processor in the user device 104. The client performs the processing required at the user device 104 in order for the user device 104 to transmit and receive data over the communication system 100.

The user device 110 corresponds to the user device 104 and executes, on a local processor, a communication client which corresponds to the communication client executed at the user device 104. The client at the user device 110 performs the processing required to allow the user 108 to communicate over the network 106 in the same way that the client at the user device 104 performs the processing required to allow the user 102 to communicate over the network 106. The

user devices

104 and 110 are endpoints in the communication system 100. FIG. 1 shows only two users (102 and 108) and two user devices (104 and 110) for clarity, but many more users and user devices may be included in the communication system 100, and may communicate over the communication system 100 using respective communication clients executed on the respective user devices.

FIG. 2 illustrates a detailed view of the user device 104 on which is executed a communication client instance 206 for communicating over the communication system 100. The user device 104 comprises a central processing unit (“CPU”) or “processing module” 202, to which is connected: output devices such as a display 208, which may be implemented as a touch-screen, and a speaker (or “loudspeaker”) 210 for outputting audio signals; input devices such as a microphone 212 for receiving audio signals, a camera 216 for receiving image data, and a keypad 218; a memory 214 for storing data; and a network interface 220 such as a modem for communication with the network 106. The user device 104 may comprise other elements than those shown in FIG. 2. The display 208, speaker 210, microphone 212, memory 214, camera 216, keypad 218 and network interface 220 may be integrated into the user device 104 as shown in FIG. 2. In alternative user devices one or more of the display 208, speaker 210, microphone 212, memory 214, camera 216, keypad 218 and network interface 220 may not be integrated into the user device 104 and may be connected to the CPU 202 via respective interfaces. One example of such an interface is a USB interface. If the connection of the user device 104 to the network 106 via the network interface 220 is a wireless connection then the network interface 220 may include an antenna for wirelessly transmitting signals to the network 106 and wirelessly receiving signals from the network 106.

FIG. 2 also illustrates an operating system (“OS”) 204 executed on the CPU 202. Running on top of the OS 204 is the software of the client instance 206 of the communication system 100. The operating system 204 manages the hardware resources of the computer and handles data being transmitted to and from the network 106 via the network interface 220. The client 206 communicates with the operating system 204 and manages the connections over the communication system. The client 206 has a client user interface which is used to present information to the user 102 and to receive information from the user 102. In this way, the client 206 performs the processing required to allow the user 102 to communicate over the communication system 100.

In acoustic echo cancellation the aim is to remove the echo signal s(t) in the microphone signal y(t) originating from the loudspeaker signal x(t). This should be done as exact as possible and as non-obtrusively as possible in order to have as little impact on the perception of any near-end signal v(t). The microphone signal can be written as y(t)=s(t)+v(t). The echo signal is a function of the loudspeaker signal as s(t)=F(x(t)).

There are two main ways to achieve the above, one being echo subtraction and the other being echo suppression. Often these two approaches are combined.

In both echo suppression and echo subtraction, a model {circumflex over (F)}(x(t)) is used to estimate the echo, or some properties of the echo such as the echo power, in the microphone signal.

A common choice in echo cancellation is to use a stochastic gradient algorithm for updating the model. A multitude of adaptation rate selection schemes have been proposed in the past with the majority of them suffering from the fact that they require the model to be roughly known in order to choose the adaptation speed. Furthermore, many of them require the accuracy of the model parameter estimates (filter coefficient estimates) to be known.

With reference to FIGS. 3 and 4 there is now described a method of removing echo. FIG. 3 is a functional diagram of a part of the user device 104 showing how an echo removal process is implemented.

As shown in FIG. 3, the user device 104 comprises the speaker 210, the microphone 212, a modelling module 302, and an echo removal module 314. The modelling module 302 comprises a filter module 304, a model accuracy determination module 306, and a convergence parameter selection module 308. The echo removal module 314 is described with reference to FIG. 3 as an echo suppression module 314.

FIG. 4 is a flow chart for the process of suppressing echo.

A signal x(t) to be output from the speaker 210 is coupled to an input of the speaker 210. It should be noted that in the embodiments described herein there is just one speaker (indicated by reference numeral 210 in the figures) but in other embodiments there may be more than one speaker to which the signal to be outputted is coupled (for outputting therefrom). Similarly, in the embodiments described herein there is just one microphone (indicated by reference numeral 212 in the figures) but in other embodiments there may be more than one microphone which receive audio signals from the surrounding environment. The signal to be output from the speaker 210 is also coupled to the modelling module 302. In particular, the signal to be output from the speaker 210 is coupled to a first input of the filter module 304. An output of the microphone 212 is coupled to the modelling module 302. In particular, the output of the microphone 212 is coupled to a second input of the filter module 304 and to a first input of the model accuracy determination module 306. The output of the microphone 212 is also coupled to a first input of the echo suppression module 314. An output of the modelling module 302 is coupled to a second input of the echo suppression module 314. In particular the output of the filter module 304 is coupled to the second input of the echo suppression module 314. An output of the echo suppression module 314 is used to provide the received signal (with echo suppression having been applied) for further processing in the user device 104. The output of the filter module 304 is also coupled to a second input of the model accuracy determination module 306. An output of the model accuracy determination module 306 is coupled to an input of the convergence parameter selection module 308. An output of the convergence parameter selection module 308 is coupled to a third input of the filter module 304.

In step S402 a signal is received which is to be outputted from the speaker 210. For example, the signal to be outputted may be a far-end signal that has been received at the user device 104 from the user device 110 during a call between the

users

102 and 108 over the communication system 100. Any processing that is required to be performed on the received signal (e.g. decoding using a speech codec, depacketizing, etc) is performed as is known in the art (e.g. by the client 206) to arrive at the signal x(t) which is suitable to be outputted from the speaker 210. The signal x(t) is a digital signal. At least some of the processing of the signal in the user device 104 prior to outputting the signal from the speaker 210 is performed in the digital domain. As is known in the art, a digital to analogue converter (DAC) is applied to the digital signal x(t) before playout from the loudspeaker 210. Similarly, an analogue to digital converter (ADC) is applied to the signal captured by the microphone 212 to arrive at the digital signal y(t).

In other embodiments, the signal to be outputted may be received from somewhere other than over the communication system 100 in a call. For example, the signal to be outputted may have been stored in the memory 214 and step S402 may comprise retrieving the signal from the memory 214.

In step S404 the audio signal x(t) is outputted from the speaker 210. In this way the audio signal x(t) is outputted to the user 102.

In step S406 the microphone 212 receives an audio signal. As shown in FIG. 3 the received audio signal may include a near-end signal which is a desired signal or “primary signal”. The near-end signal is the signal that the user 102 intends the microphone 212 to receive. However, the received audio signal also includes an echo signal resulting from the audio signals outputted from the speaker 210 in step S404. The received audio signal may also include noise, such as background noise. Therefore, the total received audio signal y(t) can be given by the sum of the near-end signal, the echo and the noise. The echo and the noise act as interference for the near-end signal.

The filter module 304 takes as inputs the outputted audio signal x(t) and the received audio signal y(t). In step S408, the filter module 304 is used to model the echo in the received audio signal y(t). In particular, the filter module 304 is operable to determine an estimate of the echo component in the near end signal y(t) using the outputted audio signal x(t) and the received audio signal y(t).

The echo path describes the effects of the acoustic paths traveled by the far end signal from the speaker 210 to the microphone 212. The far end signal may travel directly from the speaker 210 to the microphone 212, or it may be reflected from various surfaces in the environment of the near end terminal. The echo path traversed by the far end signal output from the speaker 210 may be regarded as a system having a frequency and a phase response which may vary over time.

In order to remove the acoustic echo s(t) from the signal y(t) recorded at the near-end microphone 212 it is necessary to estimate how the echo path changes the desired far-end speaker output signal x(t) to an undesired echo component in the input signal.

For an approximately linear echo path, the echo path h(t) describes how the echo in the received audio signal y(t) relates to the audio signal x(t) output from the speaker 210, e.g. according to the equation: s(t)=Σ_n=0 ^N ^trueh_n(t)x(t−n), where s(t) is the echo in the received audio signal y(t), N_trueis a sufficiently large number to cover the non-negligible parts of the impulse response (theoretically N_trueis infinite), and h_n(t) are the coefficients of the impulse response describing the echo path h(t). The echo path h(t) may vary in both time and frequency and may be referred to herein as h(t) or h(t,f). The echo path h(t) may depend upon (i) the current environmental conditions surrounding the speaker 210 and the microphone 212 (e.g. whether there are any physical obstructions to the passage of the audio signal from the speaker 210 to the microphone 212, the air pressure, temperature, wind, etc), and (ii) characteristics of the speaker 210 and/or the microphone 212 which may alter the signal as it is outputted and/or received.

The filter module 304 models the echo path h(t) of the echo in the received audio signal y(t) by determining a weighted sum of the current and a finite number (N) of previous values of the outputted audio signal x(t). The filter module 304 therefore implements an Nth order filter which has a finite length (in time) over which it considers the values of the outputted audio signal x(t) in determining the estimate of the echo path ĥ(t). In this way, the filter module 304 dynamically adapts the filter estimate of the echo path ĥ(t). The operation is described by the following equation, which defines the echo in the received audio signal y(t) in terms of the outputted audio signal x(t): ŝ₁(t)=Σ_n=0 ^Nĥ_n(t)x(t−n). Therefore N+1 samples of the outputted audio signal x(t) are used, with a respective N+1 weights ĥ_n(t). The set of N+1 weights ĥ_n(t) is referred to herein simply as the estimate of the echo path ĥ(t). In other words the estimate of the echo path ĥ(t) is a vector having N+1 values where the filter module 304 implements an Nth order filter, taking N+1 values (e.g. N+1 frames) of the signal x(t) into account.

It can be appreciated that it is easier to adapt the filter estimate of the echo path ĥ(t) when the echo is a dominant part of the received audio signal, that is when y(t)≈s(t). However, it may be possible to adapt the filter estimate of the echo path ĥ(t) even when the echo is not a dominant part of the received audio signal y(t) if the echo s(t) is independent of the other signal components of y(t).

It will be appreciated by one skilled in the art that the estimate of the echo path ĥ(t) does not need to be explicitly calculated, but could be represented by means of filter coefficients obtained from stochastic gradient algorithms such as Least Mean Squares (LMS), Normalized Least Mean Squares (NLMS), Fast Affine Projection (FAP) and Recursive Least Squares (RLS).

The filter module 304 comprises an adaptation algorithm component 310. The adaptation algorithm component 310 of the filter module 304 executes a stochastic gradient algorithm to identify the coefficients of the filter module 304 that minimises an error signal e(t).

Updated filter coefficients for the filter module 304 are generated in response to the error signal e(t), the input signal x(t) and the previous filter coefficients.

The adaptation algorithm component 310 of the filter module 304 operates in a time recursive manner. This means it does not instantaneously adapt to changes in the system, instead the algorithm iteratively converges to an approximation of the system over a finite time interval. Regardless of the particular algorithm used, the filter coefficients of the filter module 304 are updated with each iteration of the algorithm, thus the coefficients of the filter module 302 are continually updated over time regardless of the signal conditions at hand.

The filter coefficients of the filter module 304 filter the far end signal x(t) to generate an estimate of the echo component in the near end signal y(t). Whilst the above description refers to the use of a time domain FIR model of the echo path to estimate the echo component in the near end signal y(t) it will be appreciated by those skilled in the art that this is just an example and not limiting in any way. That is, the first filter module 304 may operate to determine an estimate of the echo path ĥ(t) and thus an estimate ŝ₁(t) of the echo component in the near end signal y(t) in the time domain or in the frequency domain.

The error signal e(t) is obtained by a subtractor 312 which subtracts the estimate of the echo component in the near end signal y(t) from the near end signal y(t) and supplies the error signal e(t) to the filter module 304. It will be appreciated that it is desirable for the error signal e(t) to be small. For example, when there is no near end signal v(t) in the microphone signal, ideally the error signal is equal to zero.

Stochastic gradient algorithms have a convergence parameter in the form of a step-size for the update of the model parameters. This can in some applications be chosen as fixed but in many cases better performance is achieved if it is chosen in a signal-dependent manner. The step-size controls the sensitivity of the updating to the noise in the microphone signal y(t). If it is chosen to be small, the update speed is slow but is less insensitive to the noise, but if it is chosen to be large the update speed is instead rapid but more sensitive to the noise. The reference to “update speed” or “adaptation speed” used herein is used to refer to how quickly the model is able to adapt to the signal conditions at hand in the system. That is, using a smaller step-size will result in a smaller eventual error signal e(t), however convergence to an approximation of the system will be slower due the greater number of iteration steps required (slower convergence rate), and using a larger step-size will result in a larger eventual error signal e(t), however convergence to an approximation of the system will be quicker due the fewer number of iteration steps required (faster convergence rate).

In order to achieve estimates of very high accuracy the step-size needs to be small in order to avoid overshooting the true estimates due to too high step size.

In some embodiments the estimate of the echo component is passed from the filter module 304 to the echo suppression module 314 (as shown in FIG. 3). In these embodiments the estimate of the echo component is also supplied to the model accuracy determination module 306.

In other embodiments, in step S408 the estimate of the echo component is passed from the filter module 304 to a power estimating module (not shown in FIG. 3). The power estimating module estimates the echo power in the received audio signal based on the filter estimate (determined by the filter module 304) and the far end signal x(t). There are many ways to do this that are known to persons skilled in the art and the scope of this disclosure is not limited to any particular method of determining an echo power estimate. The power estimating module is arranged to output its corresponding echo power estimate to the echo suppression module 314.

The echo suppression module 314 takes as an input an estimate of the echo component output from the first filter module 304 or an echo power estimate output from the first power estimating module, and in step S410 uses this input to apply echo suppression to the received audio signal y(t), thereby suppressing the echo in the received audio signal. The echo suppression performed at step S410 is described later.

The model accuracy determination module 306 takes as an input a first estimate ŝ₁(t). The first estimate ŝ₁(t) may be an estimate of the echo component output from the filter module 304. In one embodiment, the estimation error (y-s) output from subtractor 312 is supplied to a power estimating module (not shown in FIG. 3). In this embodiment the first estimate ŝ₁(t) may be the power of the estimation error (y-s) output from the power estimating module. The model accuracy determination module 306 also takes as an input the received audio signal y(t).

In step S412, the model accuracy determination module 306 is operable to determine the accuracy of the estimate ŝ₁(t).

One metric used for measuring the accuracy of the filter module 304 is the echo return loss enhancement (ERLE), the ERLE metric is defined as:

E R L E_{k} (t) = \frac{{E [y (t)]}^{2}}{{E [{\hat{s}}_{k} (t) - y (t)]}^{2}}

The ERLE metric may be measured in decibels (dB), according to the following equation (assuming that a base 10 logarithm is used i.e., log 10):

E R L E_{k} (t) = 10 \log \frac{{E [y (t)]}^{2}}{{E [{\hat{s}}_{k} (t) - y (t)]}^{2}}

In the above equations, E[ ] is the expectation operator. The ERLE measure can, and typically is, applied to non-stationary signals. Therefore, in practice the expectation values are evaluated using short-time average values:

E R L E_{k} (t) = \frac{\sum_{k = 0}^{K - 1} y^{2} (t - k)}{\sum_{k = 0}^{K - 1} {(\hat{s} (t - k) - y (t - k))}^{2}}

The ERLE metric is a measure of the ability of the model estimate to model the microphone signal y(t). The ERLE metric is limited in the sense that it only assesses the model accuracy when the microphone signal mainly consists of echo. If that is not the case, the ERLE measurement may be low even though the model is accurate. However, if the ERLE measurement is high, it can only be due to the model being accurate.

The accuracy value of the estimate ŝ₁(t) may be determined periodically. For example, the accuracy value of the estimate ŝ₁(t) may be averaged over a predetermined number of samples of the audio signal x(t) and the received audio signal y(t) in a given time period to arrive at the accuracy value. That is, the accuracy value may be determined for each frame of the received audio signal y(t) however this is merely an example, and the respective accuracy value may be determined less or more often than for each frame.

Once the accuracy of the estimate ŝ₁(t) has been determined in step S412, the process proceeds to step S414.

In step S414, the convergence parameter selection module 308 determines a value for a convergence parameter (step size) used in an algorithm executed by the adaptation algorithm component 310 of the filter module 304 based on the accuracy of the estimate ŝ₁(t) determined in step S412. The convergence parameter is algorithm dependent, but can be quantified for a certain algorithm. Taking the NLMS algorithm as an example, a convergence parameter of 0.5 would typically provide a reasonably fast step size (fast update speed) and a convergence parameter of 0.01 would typically give a very small step size (slow update speed).

In particular, the convergence parameter selection module 308 selects the convergence parameter to control the adaptation speed of the filer module 304 as a non-increasing function of the echo return loss enhancement measurement made in step S412.

In step S414, the echo return loss enhancement measurement made in step S412 may be compared to a threshold value, and the convergence parameter selection module 308 adjusts the convergence parameter based on this comparison. The threshold value may be a predetermined threshold, for example, 10 dB. However it will be appreciated that this value for the predetermined threshold is merely an example and it may be set differently according to system requirements.

That is, if the echo return loss enhancement measurement taken in step S412 is less than the threshold value (i.e. when the ERLE value is low) the convergence parameter selection module 308 selects a convergence parameter value to increase the adaptation speed of the filer module 304.

The decision to increase the adaptation speed of the filer module 304 may be made only if the echo return loss enhancement measurement has been less than the threshold value over a predetermined period of time i.e. when the echo return loss enhancement measurement has been less than the threshold value for the whole duration of the predetermined period of time. Other related schemes could also be used, for example the decision to increase the adaptation speed of the filer module 304 may be made if the echo return loss enhancement measurement has been less than the threshold value for a predetermined proportion of a predetermined period of time. By setting the predetermined proportion accordingly, in one example, the adaptation speed is increased when the ERLE has been less than the threshold almost all of the predetermined period of time, apart from at a few samples where it has exceeded the threshold.

Depending on the algorithm executed by the adaptation algorithm component 310, the increase of the adaptation speed of the filer module 304 may be implemented in a number of ways. That is, the increase of the adaptation speed of the filer module 304 may be implemented by decreasing the convergence parameter by a predetermined amount or to a predetermined value. Alternatively, the increase of the adaptation speed of the filer module 304 may be implemented by increasing the convergence parameter by a predetermined amount or to a predetermined value.

If the echo return loss enhancement measurement taken in step S412 is equal to or greater than the threshold value (i.e. when the ERLE value is high) the convergence parameter selection module 308 selects a convergence parameter value to decrease the adaptation speed of the filer module 304.

The decision to decrease the adaptation speed of the filer module 304 may be made only if the echo return loss enhancement measurement has been equal to or greater than the threshold value over a predetermined period of time i.e. when the echo return loss enhancement measurement has been equal to or greater than the threshold value for the whole duration of the predetermined period of time. Other related schemes could also be used, for example the decision to decrease the adaptation speed of the filer module 304 may be made if the echo return loss enhancement measurement has been equal to or greater than the threshold value for a predetermined proportion of a predetermined period of time. By setting the predetermined proportion accordingly, in one example, the adaptation speed is decreased when the ERLE has been equal to or greater than the threshold almost all of the predetermined period of time, apart from at a few samples where it has dropped below the threshold.

Depending on the algorithm executed by the adaptation algorithm component 310, the decrease of the adaptation speed of the filer module 304 may be implemented in a number of ways. That is, the decrease of the adaptation speed of the filer module 304 may be implemented by decreasing the convergence parameter by a predetermined amount or to a predetermined value. Alternatively, the decrease of the adaptation speed of the filer module 304 may be implemented by increasing the convergence parameter by a predetermined amount or to a predetermined value.

In embodiments described above, more than one threshold may be used. Taking the example implementation whereby the increase/decrease of the adaptation speed of the filer module 304 is implemented by increasing/decreasing the convergence parameter to a predetermined value. An example could be that on the condition that the ERLE≦2 dB, the convergence parameter is set to 0.5; on the condition that 2 dB<ERLE<4 dB, the convergence parameter is set to 0.2; and on the condition that the ERLE≧4 dB, the convergence parameter is set to 0.05. It will be appreciated that these specific values for the convergence parameter are merely used for illustration purposes. The values chooses for the convergence parameter are application dependent.

In embodiments described above the convergence parameter selection module 308 may implement hysteresis control to vary the convergence parameter more slowly. For example an upper threshold and lower threshold may be selected centred on the threshold value, and the decreasing of the adaptation speed of the filer module 304 described above may only be implemented when the echo return loss enhancement measurement taken in step S412 is equal to or greater than the upper threshold, and the increasing of the adaptation speed of the filer module 304 described above may only be implemented when the echo return loss enhancement measurement taken in step S412 is less than the lower threshold. This prevents rapid variations to the convergence parameter as the echo return loss enhancement measurement drifts around the threshold value.

Since the echo return loss enhancement measure has the property that the model accuracy is always high when the echo return loss enhancement measurement is high it may be used to slow down the adaptation speed when the echo return loss enhancement measurement is high in order to achieve increasingly accurate estimates, and increase the adaptation speed when the echo return loss enhancement measurement is low in order to quickly track changes in the model parameters.

Embodiments choose the adaptation speed as a function of the accuracy of the model, but this accuracy is implicitly estimated as part of the selection scheme.

The embodiments described above ensures that fast adaptation is achieved when the accuracy of the model is unknown (via the high updating speed when the echo return loss enhancement measurement is low), and that increasingly accurate estimates are achieved when the model is known to be accurate (via decreasing the updating speed when the echo return loss enhancement measurement is high).

The adaptation scheme in embodiments described above is particularly useful in updating schemes where the model is continuously updated and where it is of primary interest to ensure that the model estimates at some (known) points in time are highly accurate, rather than that the model estimates are fairly accurate all the time.

Example architectures in which a model is continuously updated and where the above adaptation scheme may be used, are now described.

In one example architecture, the modelling module is configured to determine a first model estimate of the echo in the received audio signal using a first model based on the outputted audio signal x(t) and the received audio signal y(t), and determine a second model estimate of the echo in the received audio signal using a second model based on the outputted audio signal. The first model is continually updated over time. The modelling module is configured to determine a first accuracy value of the first model according to a model accuracy measure, determine a second accuracy value of the second model according to the model accuracy measure, and determine if the first model is more accurate than the second model based on a comparison of the first accuracy value and the second accuracy value and selectively update the second model based on the comparison. That is, the second model is updated if the first model is more accurate than the second model, and the second model is not updated if the first model is not more accurate than the second model. In this architecture, an echo removal module is configured to use only the second model estimate of the echo to remove the echo in the received audio signal. Thus two adaptive models are used to produce model estimates in a shadow manner. In this architecture the step size adjustment scheme described above may be used to vary the step size of the algorithm used in updating the first model. That is, the echo return loss enhancement measure may be used to slow down the adaptation speed of the first model when the echo return loss enhancement measurement of the first model is high in order to achieve increasingly accurate estimates, and increase the adaptation speed when the echo return loss enhancement measurement is low.

In another example architecture, the modelling module is configured to model an echo path of the echo in the received audio signal using a first model based on the outputted audio signal and the received audio signal to determine a first model estimate of the echo in the received audio signal, use the first model estimate to determine a first performance value according to a performance metric, compare the first performance value with a threshold value and determine if the echo path can be deemed linear based on the comparison. The modelling module is further configured to selectively model the echo path of the echo in the received audio signal using a second model based on the outputted audio signal and the received audio signal, based on said comparison, to determine a second model estimate of the echo. In this architecture, an echo removal module configured to selectively use the first model estimate or the second model estimate of the echo to remove the echo in the received audio signal based on the comparison. That is, the first model estimate of the echo is used to remove the echo in the received audio signal if it is determined that the echo path can be deemed linear, and the second model estimate of the echo is used to remove the echo suppression in the received audio signal, if it is determined that the echo path cannot be deemed linear. In this architecture the first model is continually updated over time. Thus, the step size adjustment scheme described above may be used to vary the step size of the algorithm used in updating the first model. That is, the echo return loss enhancement measure may be used to slow down the adaptation speed of the first model when the echo return loss enhancement measurement of the first model is high in order to achieve increasingly accurate estimates, and increase the adaptation speed when the echo return loss enhancement measurement is low.

The echo suppression performed at step S410 is now described. The purpose of the echo suppressor 314 is to suppress the loudspeaker echo present in the microphone signal, e.g. in a VoIP client, to a level sufficiently low for it not to be noticeable/disturbing in the presence of the near-end sounds (non-echo sounds) picked up by the microphone 212. The echo suppression module 314 is designed to apply signal dependent suppression that varies both over time and frequency to the received audio signal y(t). Echo suppression methods are known in the art. Furthermore, the echo suppression method applied by the echo suppression module 314 may be implemented in different ways. As such, the exact details of the echo suppression method are therefore not described in detail herein.

The echo suppression module 314 outputs the received signal, with the echo having been suppressed, for further processing at the user device 104. For example, the signal output from the echo suppression module 314 may be processed by the client 206 (e.g. encoded and packetized) and then transmitted over the network 106 to the user device 110 in a call between the

users

102 and 108. Additionally or alternatively, the signal output from the echo suppression module 314 may be used for other purposes by the user device 104, e.g. the signal may be stored in the memory 214 or used as an input to an application which is executing at the user device 104.

The filter module 304 may utilise any filter implementing a stochastic gradient algorithm. The filer module 304 may be a linear to model the echo path of the echo in the received audio signal (e.g. a Finite Impulse Response (FIR) filter or an Infinite impulse Response (IIR) filter) or alternatively may utilise a non-linear filter.

In the embodiments described above, the echo removal is implemented in a VoIP system (e.g. the received audio signal may include speech of the user 102 for transmission to the user device 110 during a call between the

users

102 and 108 over the communication system 100). However, the echo removal methods described herein can be applied in any suitable system in which echo removal is to be applied.

In the embodiments described above, and shown in the Figures, the echo removal module 314 implements echo suppression.

In the embodiments described above, and shown in the Figures, echo cancellation (or “echo subtraction”) is not applied to the received audio signal y(t). That is, there is no echo cancellation module in the user device 104 and the echo suppression is applied to the received audio signal y(t) without a prior step of applying echo cancellation to the received audio signal y(t).

However, in other embodiments, echo cancellation may be applied, by an echo cancellation module, to the received audio signal y(t). In particular, the echo suppression applied by the echo suppression module 314 may be applied downstream of (i.e. after) the echo cancellation in the processing of the received audio signal y(t). The echo cancellation module would subtract an estimate of the echo signal from the received audio signal, but due to inaccuracies in the estimate of the echo signal, a residual echo would most-likely remain in the received audio signal. It is the residual echo that would then be suppressed by the echo suppression module 314. This echo suppression could be applied in the same way as described herein in the embodiments in which no echo cancellation is applied. If echo subtraction is used, the effect of it can be taken into account in the echo suppression.

In other embodiments, the echo removal module 314 implements echo cancellation (at step S410). That is, the echo removal module 314 is arranged to subtract an estimate of the echo signal ŝ₁(t) from the received audio signal y(t).

The methods described herein may be implemented as the sole mechanism for adjusting the step size of the algorithm used in updating the adaptive model. Alternatively the methods described herein may be implemented as one component of a step size adjustment scheme wherein other considerations are made when adjusting the step size of the algorithm.

The methods described herein may be implemented by executing a computer program product (e.g. the client 206) at the user device 104. That is, a computer program product may be configured to remove echo in the received audio signal y(t), wherein the computer program product is embodied on a computer-readable storage medium (e.g. stored in the memory 214) and configured so as when executed on the CPU 202 to perform the operations of any of the methods described herein.

Generally, any of the functions described herein (e.g. the functional modules shown in FIG. 3 and the functional steps shown in FIG. 4) can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), or a combination of these implementations. The modules and steps shown separately in FIGS. 3 and 4 may or may not be implemented as separate modules or steps. For example, the filter module 304 may perform the function of the model accuracy determination module 306, the convergence parameter selection module 308 and of the subtractor 312. The terms “module,” “functionality,” “component” and “logic” as used herein generally represent software, firmware, hardware, or a combination thereof. In the case of a software implementation, the module, functionality, or logic represents program code that performs specified tasks when executed on a processor (e.g. CPU or CPUs). The program code can be stored in one or more computer readable memory devices. The features of the techniques described herein are platform-independent, meaning that the techniques may be implemented on a variety of commercial computing platforms having a variety of processors. For example, the user devices may also include an entity (e.g. software) that causes hardware of the user devices to perform operations, e.g., processors functional blocks, and so on. For example, the user devices may include a computer-readable medium that may be configured to maintain instructions that cause the user devices, and more particularly the operating system and associated hardware of the user devices to perform operations. Thus, the instructions function to configure the operating system and associated hardware to perform the operations and in this way result in transformation of the operating system and associated hardware to perform functions. The instructions may be provided by the computer-readable medium to the user devices through a variety of different configurations.

One such configuration of a computer-readable medium is signal bearing medium and thus is configured to transmit the instructions (e.g. as a carrier wave) to the computing device, such as via a network. The computer-readable medium may also be configured as a computer-readable storage medium and thus is not a signal bearing medium. Examples of a computer-readable storage medium include a random-access memory (RAM), read-only memory (ROM), an optical disc, flash memory, hard disk memory, and other memory devices that may us magnetic, optical, and other techniques to store instructions and other data.

Although the subject matter has been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Claims

The invention claimed is:

1. A method of removing echo, the method comprising:

outputting an output audio signal;

receiving an input audio signal, wherein the received input audio signal includes echo resulting from said output audio signal;

determining a first adaptive model estimate of the echo in the received input audio signal based on the outputted output audio signal and the received input audio signal;

determining a performance value of the first adaptive model according to a performance metric;

comparing the performance value with a threshold value;

determining a second adaptive model estimate of the echo in the received input audio signal based on the outputted output audio signal, the received input audio signal, and further based on the comparison of the performance value with the threshold value;

determining, based on the comparison of the performance value with the threshold value and further based on a relation between echo in the received input audio signal and the outputted output audio signal, whether an echo path is deemed linear; and

in an event that the echo path is deemed linear:

using the first model estimate of the echo to remove the echo in the received input audio signal and updating a convergence parameter to control an adaptation speed of the first adaptive model; and

in an event that the echo path is not deemed linear:

using the second model estimate of the echo to remove the echo in the received input audio signal and updating the convergence parameter to control the adaptation speed of the first adaptive model.

2. The method of claim 1, wherein if the performance value is equal to or greater than the threshold value, said updating comprises selecting a convergence parameter value to decrease the adaptation speed of the first adaptive model.

3. The method of claim 1, wherein if the performance value is equal to or greater than the threshold value for a predetermined period of time, said updating comprises selecting a convergence parameter value to decrease the adaptation speed of the first adaptive model.

4. The method of claim 1, wherein if the performance value is equal to or greater than the threshold value for a predetermined proportion of a predetermined period of time, said updating comprises selecting a convergence parameter value to decrease the adaptation speed of the first adaptive model.

5. The method of claim 1, wherein if the performance value is less than the threshold value, said updating comprises selecting a convergence parameter value to increase the adaptation speed of the first adaptive model.

6. The method of claim 1, wherein if the performance value is less than the predetermined threshold value for a predetermined period of time, said updating comprises selecting a convergence parameter value to increase the adaptation speed of the first adaptive model.

7. The method of claim 1, wherein if the performance value is less than the predetermined threshold value for a predetermined proportion of a predetermined period of time, said updating comprises selecting a convergence parameter value to increase the adaptation speed of the first adaptive model.

8. The method of claim 1, wherein said updating comprises:

decreasing the convergence parameter by a predetermined amount or to a predetermined value; or

increasing the convergence parameter by a predetermined amount or to a predetermined value.

9. The method of claim 1, wherein the convergence parameter is a step size parameter.

10. The method of claim 1, wherein using the first adaptive model estimate of the echo to remove the echo in the received input audio signal comprises using the first adaptive model estimate of the echo to apply echo suppression to the received input audio signal, thereby suppressing the echo in received input audio signal.

11. The method of claim 1 wherein using an adaptive model estimate of the echo to remove the echo in the received input audio signal comprises using the first or second adaptive model estimate of the echo to apply echo cancellation to the received input audio signal, wherein the first or second adaptive model estimate of the echo is subtracted from the received input audio signal.

12. The method of claim 1, wherein the method is performed at a user device for use in a communication event, and wherein the received input audio signal comprises speech of a user for transmission from the user device in the communication event.

13. The method of claim 12, wherein the communication event is a voice-over-internet-protocol (VoIP) call.

14. The method of claim 13, wherein the outputted output audio signal comprises far-end speech signals of the VoIP call which are included in the echo in the received input audio signal.

15. The method of claim 1, wherein the performance metric comprises an echo return loss enhancement metric.

16. A device comprising:

audio output apparatus configured to output an output audio signal;

audio input apparatus configured to receive an input audio signal, wherein the received input audio signal includes an echo resulting from said outputted output audio signal;

a modelling module configured to determine a first adaptive model estimate of the echo in the received input audio signal based on the outputted output audio signal and the received input audio signal, the modelling module further configured to:

determine a performance value of the first adaptive model according to a performance metric;

compare the performance value to a threshold;

determine a second adaptive model estimate of the echo in the received input audio signal based on the outputted output audio signal, the received input audio signal, and further based on the comparison of the performance value with the threshold value;

determine, based on the comparison of the performance value with the threshold value and further based on a relation between echo in the received input audio signal and the outputted output audio signal, whether an echo path is deemed linear; and

in an event that the echo path is deemed linear:

in an event that the echo path is not deemed linear:

17. The device of claim 16, wherein the audio output apparatus comprises a speaker configured to output the outputted output audio signal, and wherein the audio input apparatus comprises a microphone configured to receive the received input audio signal.

18. A computer program product configured to remove echo in a received audio signal, said echo resulting from an outputted audio signal, the computer program product being embodied on a computer-readable storage medium and configured so as when executed on a processor to:

determine a first adaptive model estimate of the echo in the received audio signal based on the outputted audio signal and the received audio signal;

compare the performance value with a threshold value;

determine, based on the comparison of the performance value with the threshold value and further based on a relation between echo in the received input audio signal relates to the outputted output audio signal, whether an echo path is deemed linear; and

in an event that the echo path is deemed linear:

in an event that the echo path is not deemed linear:

19. The computer program product of claim 18, wherein if the performance value is equal to or greater than the threshold value for a predetermined proportion of a predetermined period of time, said updating comprises selecting a convergence parameter value to decrease the adaptation speed of the first adaptive model.

20. The computer program product of claim 18, wherein if the performance value is less than the threshold value, said updating comprises selecting a convergence parameter value to increase the adaptation speed of the first adaptive model.